Abstract: Partial discharge (PD) detection is an important technique for insulation condition evaluation of gas insulated switchgear (GIS)/gas insulated transmission line(GIL). Ultra-high frequency (UHF) and acoustic emission (AE) methods are commonly used in on-site PD detection. However, due to the complex and serious interference and the significant propagation attenuation of PD signals, the signal-to-noise ratio (SNR) of PD signals is extremely low, and even the PD signals are completely submerged in noise, which makes it difficult for PD diagnosis and localization. Therefore, in this paper, a denoising method for PD UHF and AE signals based on coherent averaging was proposed. Compared with the wavelet method and singular value decomposition (SVD) method, the proposed method has lower mean square error (MSE), higher normalized correlation coefficient (NCC) and reduction in noise level (RNL), and it does not require complex parameter selection. Then, on-site PD detection was conducted at a hydropower plant, and denoising and localization analysis on UHF and AE signals was performed. The results show that, by using the proposed method, the noise of the PD signal can be reduced from over ten mV to below 1 mV, and the PD pulse submerged in noise can be effectively extracted. Especially, with extremely low SNR, the proposed method can get good denoising performance, while the traditional methods are ineffective. Based on the denoised signals, precise defect localization can be achieved, which verifies the effectiveness of the proposed denoising method. The results of this paper can effectively improve the effectiveness of on-site PD detection, providing important support for GIS/GIL defect detection and localization.